Multi-Resolution Aitchison Geometry Image Denoising for Low-Light Photography.

Area of Science:

• Image processing
• Computer vision
• Statistical modeling

Background:

• Low-photon imaging is susceptible to shot noise, often modeled as a Poisson distribution.
• Traditional denoising methods may struggle to preserve crucial image details in this regime.

Purpose of the Study:

• To develop a robust image denoising technique for low-photon imaging.
• To enhance the preservation of image contrast details relevant to human perception.

Main Methods:

• Utilized Bayesian estimation with a Poisson likelihood function.
• Applied statistical compositional data analysis and reinterpreted Aitchison geometry.
• Developed a denoising technique using an approximate conjugate prior for latent variables.

Main Results:

• Demonstrated a strong match between Poisson likelihood and Bayesian estimation of difference-log-contrast.
• Showcased wavelet-like properties of difference-log-contrast, aligning with the human visual system.
• Achieved robust denoising resilient to illumination variations.

Conclusions:

• The proposed method effectively denoises images in the low-photon regime.
• The technique preserves image contrast details, improving perceptual quality.
• Difference-log-contrast offers a robust feature for image analysis and denoising.